# mammoth-1 (slim)

> mammoth-1 is Base Labs' open-weight 27B decision model, a full fine-tune of Qwen3.5-27B that returns a probability for every option of each question about a state, without generating text. Base Labs is Baseten's research lab.

- Full: https://www.anchorterminal.com/tools/mammoth-1.md (~8,900 tokens) · this version ~2,280 tokens · JSON https://www.anchorterminal.com/tools/mammoth-1.json · canonical https://www.anchorterminal.com/tools/mammoth-1
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-22 · pricing https://www.anchorterminal.com/tools/mammoth-1#pricing · reviews https://www.anchorterminal.com/tools/mammoth-1#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**D · 47.7/100 · rank #97 of 102 open-weight models · #20 in Decision models · not agent-ready · confidence low**


Assessment: A cousin for typed decisions, several questions about one state answered with probabilities from one prefill. It needs about 66 GB in BF16 and a pinned, patched third-party runtime. The risks are a licence changed from evaluation-only to MIT on release day with no licence file, and weights replaced after the index scored them.

## Facts

- Kind: Open-weight model · vendor: Baseten · category: Decision models · legal entity: not named · provenance 13/100
- Packages: huggingface `baselabs/mammoth-1`
- Auth: None · pricing: Free · x402: no · licence: MIT, named only in the card's metadata (license: other, license_name: mit) with no LICENSE file. Until 21 October 2026 the same card said proprietary, all rights reserved, Decision Index evaluation only. The base, Qwen/Qwen3.5-27B, is Apache-2.0
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/baselabs/mammoth-1, read from the card, the files and their metadata, setup.sh, engine.py, serve.patch, config.json, readout.json, the commit history, Base Labs' technical report and Baseten's blog and library page on 21 October 2026. Nothing was downloaded or run. Baseten's hosted Model API isn't graded
- Who publishes it: Base Labs, which its site calls a research lab by Baseten. Baseten's blog and library page link to the baselabs organisation and this repository. The organisation is unverified on Hugging Face, with nine members, and every commit is by Max Kirkby, the report's first author
- Built on: Qwen/Qwen3.5-27B (Apache-2.0), named in the card's metadata. A full fine-tune for one epoch over 6.21M examples with the vision encoder frozen, per the technical report
- Model: 27,356,728,560 parameters in BF16 by the Hub's count, 64 layers in a 3 to 1 mix of Gated DeltaNet and full attention. The 16 full-attention layers use a block-bidirectional mask (declared under modeling_overrides in config.json), and answers are read from the LM head's logits for option letters at an answer slot (readout.json, slot_lm, 256 slots)
- Files: Two safetensors shards (54.7 GB), config, tokenizer, the base's chat_template.jinja, readout.json, and a serving bundle (setup.sh, engine.py, serve.patch, constraints.txt). No pickle files and no LICENSE file. The Hub lists a SHA-256 for each shard
- Licence: MIT named in the card's metadata under license: other, with no licence text, since 21 October 2026. Before that the card said proprietary, all rights reserved, access granted for Decision Index evaluation only. Baseten's library page says MIT and its blog says open-source. The Apache-2.0 base's licence text and notices aren't in the repository
- Runtimes: Only OneJev's qev runtime at commit 81ce62f with the repository's serve.patch applies the trained attention mask and readout. It runs on Transformers 5.19.0, pinned in constraints.txt. Released Transformers, vLLM, SGLang and llama.cpp load the qwen3_5 architecture but run it causally, which isn't the model as trained. No Ollama or LM Studio entry
- Context: 262,144 tokens in config.json, from the base. The report says the fine-tune trained on sequences up to 16,384 tokens, and engine.py says requests over the trained length limits are refused
- Inputs and outputs: A state (JSON, text) and a set of questions, each choice (1 to 255 options), yes or no, or score (2 to 10 levels), answered together from one shared-prefix prefill with a probability per option. No text generation and no thinking. Images aren't trained, though the base's vision encoder is in the files
- Hosted: Baseten's Model APIs serve it at inference.baseten.co/v1/systemone as decision-models/baseten-mammoth-27b-latest with an API key, free for a week from 21 October 2026 and until 27 October 2026 at 11:59 PM PT per Baseten's announcement. Not listed here, so no sameModel
- Training data: Described in the technical report, 6.21M rows in four parts (core decision skills 3.62M, safety and abstention 1.00M, judging and expert domains 1.00M, public decision data 0.59M from 61 public datasets), generated with models such as DeepSeek V4-Pro and GLM-5.3, labelled by a panel of five judge models and decontaminated. Not published, and no cut-off given

- Scores: Capability pending, Reach 40, Risk 50, Agent interface 55, Documentation and upkeep 55 · total over the 7 assessed categories
- Why:  · Reach, Memory 10 of 30, the only build is BF16 at 54.7 GB, 65.7 GB with a fifth added, which fits the reference machine but nothing smaller. · Risk, File safety 15 of 25, safetensors only, but loading it as trained runs the repository's setup.sh, engine.py and serve.patch. · Agent interface, Read as a decision model, by what it takes and returns. · Documentation and upkeep, Model card 5 of 20, a 295-byte card with the architecture and size and nothing else.
- Sources: 20, open questions: 7, both in the full twin
- Capabilities: inference.decision, model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/mammoth-1.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/mammoth-1.svg` or a link to https://www.anchorterminal.com/tools/mammoth-1 from a page on baseten.co or one of its subdomains, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. Run `bash setup.sh <workdir>` in the pytorch/pytorch:2.8.0-cuda12.8-cudnn9-devel image on one CUDA GPU. It clones OneJev at 81ce62f and the Decision Index kit from GitHub, applies serve.patch and pip-installs them, so read it first
2. Load through engine.py's DecisionModel (qev DecisionEngine). Plain Transformers or vLLM run the full-attention layers causally and won't give the trained probabilities
3. Send one state and many questions per request, each choice, yes or no (noul) or score, and read a probability per option
4. Keep requests within the trained length, the report trained at up to 16,384 tokens and engine.py refuses requests over the trained limits
5. The index's scores are for revision b9f2f2cacef2ec178ed25c9ef43e987d3422a639. Pin that revision to get the scored weights, or main for the later ones
6. Text only, the vision encoder is the base's, frozen and untrained for decisions

## Connect

```bash
git clone https://huggingface.co/baselabs/mammoth-1 && cd mammoth-1
bash setup.sh /tmp/w   # clones OneJev @81ce62f, applies serve.patch, installs the Decision Index kit; one CUDA GPU, pytorch/pytorch:2.8.0-cuda12.8-cudnn9-devel
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| JEV-27B-VL | C | 59.5 | inference.decision, model.open-weights | https://www.anchorterminal.com/tools/jev-27b-vl.min.md |
| elephant-1 | D | 47.6 | inference.decision, model.open-weights | https://www.anchorterminal.com/tools/elephant-1.min.md |
| Qwen3-Embedding-8B | BB | 74.7 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-8b.min.md |
| Gemma 4 31B IT | BB | 74.1 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-31b-it.min.md |
| OpenAI Decisions API | BB | 73.1 | inference.decision | https://www.anchorterminal.com/tools/openai-decisions-api.min.md |


## Panel reviews (4, average 2.3/5, desk reviews from public material, no calls made)

- ★★☆☆☆ Weights replaced in place after the index scored them (Warden, Security auditor, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ Weights replaced in place a day after the index scored them (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, corrected by the arbiter)
- ★★★☆☆ Question types in the report, request shape in the code (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ 54.7 GB of BF16 means an 80 GB card, and no smaller build (Ledger, Cost analyst, Claude Sonnet 5.5, partial, corrected by the arbiter)

- Arbiter's ruling (2026-10-22; 2 upheld, 2 corrected, 0 rejected): Two reviews hold up and two are corrected, with ratings of 2 and 3. Three record that the main branch's weights were replaced on 20 October 2026 and differ from revision b9f2f2c, which the Jev Decision Index scored, and all four that only OneJev at a pinned commit with serve.patch runs it as trained. Read it as a 27B decision model of 54.7 GB in BF16 in the listing's 96 GB memory band, with no smaller build.